Optimizing Blood Cell Component Detection with Deep Learning: Evaluating the Performance of Transfer Learning ResNet50 and Conventional CNN Models

In this study, ResNet50 was employed to extract feature representations of blood cell components based on image analysis and classify them, focusing on the application of transfer learning. The power of transfer learning, its speed, efficiency, and resource optimization compared to traditional training methods were examined and demonstrated. Microscopic image data was used, along with various techniques to enhance the model’s predictive accuracy, feature comparisons and analyses were made. Predictions for blood cell component detection and classification were then generated. The dataset used includes 17,092 original images (classified into basophils, eosinophils, neutrophils, erythroblasts, lymphocytes, monocytes, and platelets) and 85,460 augmented images. The data is publicly available under the CC BY-SA 4.0 license. The proposed method yielded an expected accuracy of 98%, demonstrating the feasibility of applying this approach to blood cell analysis and classification.